lead · github
12-Factor Agents: Production Rules for LLM Software
Humanlayer's framework applies proven deployment principles to agent-based systems. If you're shipping AI products to real customers, this is the checklist you need.
humanlayer/12-factor-agents ↗Building LLM agents feels like the Wild West right now. You can get a prototype working in an afternoon, but moving from "works on my machine" to "reliable in production" is where most builders hit a wall.
Humanlayer's 12-Factor Agents framework borrows from 12-factor app methodology (the battle-tested standard for building maintainable web services) and rewrites it for agents. The core idea: stateless design, explicit configuration, proper logging, graceful failure modes, and human oversight baked in from the start.
Why it matters: agents that work in demo often break silently in production. These 12 principles address the specific failure modes of LLM-powered systems: hallucinations propagating unchecked, state getting lost mid-task, deployment drifting between environments, and no audit trail when things go wrong.
The repo includes both principles and working code. If you're shipping an agent product or deploying AI to prod, this cuts weeks off the "make it reliable" phase.
Share kit
12-Factor Agents: Proven rules for shipping LLM products to production
Humanlayer just open-sourced a framework that applies 12-factor app principles to LLM agents. If you're moving beyond prototypes, this is the production checklist: stateless design, explicit config, structured logging, human oversight loops, and more. Code included. Worth 20 minutes if you're shipping agent products or deploying to real customers.
agents look easy until they hit production. Humanlayer just open-sourced 12-Factor Agents, a framework for building LLM systems that don't silently break in the wild. borrowed from 12-factor apps but rewritten for the failure modes that actually kill agent products: hallucinations, lost state, no audit trail. code included. https://github.com/humanlayer/12-factor-agents
The gap between "working prototype" and "production-ready agent" is where most AI builders get stuck. Humanlayer's 12-Factor Agents framework bridges that gap by adapting proven SaaS deployment principles to LLM-powered systems. The core: stateless design, explicit configuration, structured logging, human-in-the-loop oversight, and graceful degradation. If you're shipping agent products or deploying AI systems to customers, this is the operational foundation that separates one-off demos from reliable, maintainable software. Open source, code included. https://github.com/humanlayer/12-factor-agents
Been thinking about why most agent deployments fall apart in production. The gap isn't capability, it's architecture. Just found 12-factor-agents, and it's doing something overdue: taking the 12-factor app methodology (the thing that actually made web services reliable) and translating it for LLM agents. No theory. Concrete patterns for reproducibility, observability, statelessness, and configuration that separate toys from systems customers can trust. If you're shipping agents to prod, this is the blueprint.
the thing about agents is they're still built like prototypes. reproducibility is a guess. observability is a log file. state lives everywhere. 12-factor-agents ports the methodology that fixed this for web apps to llm agents. concrete patterns, not wishes. if you're putting agents in front of customers, this is the checklist you need.